Re: [patch] JNI performance improvements

[Bryce McKinlay <mckinlay at redhat dot com>]

graydon hoare wrote:

> this patch implements two JNI optimizations:
>
>  - avoidance of the malloc/free on each JNI call, when the call only
>    uses the "bottom", non-reentrant frame (this is the majority of JNI
>    traffic).
>
>  - a fast path which avoids clearing the frame when leaving it, if
>    no local references have been made. most JNI functions do not make
>    any local references.


Graydon,

Great patch! But... are you sure that caching only the bottom-most frame 
is always sufficient? Its easy to imagine cases where nested JNI calls 
will be common. For example, won't the AWT JNI code be re-entered 
frequently now that we have a single event dispatch thread? ie: 
gtk_main_iteration called from event thread via JNI, an event is 
dispatched back to application which then makes another JNI call? Seems 
that we could end up popping up and down different levels of the stack 
frequently, which will eliminate much of the benefit of this 
optimization as frames will keep getting freed and re-allocated just as 
before.

We could avoid this by maintaining a chain of bottom frames which get 
re-used. Then again, if profiling shows that this really doesn't happen 
in practice then I'm prepared to trust that.

> +// Destroy the env's reusable resources. This is called from the thread
> +// destructor "finalize_native" in natThread.cc
> +void 
> +_Jv_FreeJNIEnv (_Jv_JNIEnv *env)
> +{
> +  if (env == NULL)
> +    return;
> +
> +  if (env->bottom_locals != NULL)
> +    _Jv_Free (env->bottom_locals);
> +
> +  _Jv_Free (env);
> +}
>  
Why don't we just GC-allocate these - ie _Jv_AllocBytes() ? Then we 
wouldn't need _Jv_FreeJNIEnv.

Regards

Bryce